Part 2 - Constructs and Estimation
University of Lübeck
University of Lübeck
Trinity College Dublin
National Tsing Hua University
As researchers in HCI or in the social sciences, we study human properties (operationalized as variables) and how they are related.
From properties to variables…
… using survey items: Structural equation modeling allows us to examine the relationship between items (the indicators) and variables (the constructs).
If we research relationships between variables, we are often interested in causality. Structural equation modeling uses regression to model relationships between variables (or constructs). But regression does not imply causality.
The gold standard for examining causality in science is the randomized controlled trial. But in HCI research, we often only have observational data.
From association …
… to causation?
But not all is lost.
Causal interpretation is possible if
using causal modeling techniques, e.g., Directed Acyclic Graphs (DAGs) or Structural Equation Modeling
Further reading: The Causal Foundations of Structural Equation Modeling. [2]
Structural equation modeling then allows us to examine two relationships at the same time:
These relationships are defined in two models that are linked in the estimation process:
We know take you through the process of SEM estimation as you would go through it in your own research.
We have models to predict behavioral intention (BI).
Unified Theory of Acceptance and Use of Technology [3]
We are interested in examing UTAUT2 for Trello, specifically our RQ is:
UTAUT2 has a strong theory and evidence base. This gives us some support for causal interpretation.
From your research questions, you will derive some hypotheses.
We hypothesize that Performance Expectancey (PE), Effort Expectancy (EE), and Social Influence (SI) influence BI.
So far this is equivalent to a regular multiple regression.
But we might also hypothesize that
These hypotheses can be directly translated into code. And this code can be visualized as a DAG.
Now, we think about how we want to measure these constructs.
Luckily, for UTAUT2, there are some standardized items so that part is made easy for us.
Example - Behavioral Intention (Scale 1-7):
There is a distinction between reflective and formative measurement.
Reflective means the scale is the result of a latent construct (openness).
Formative means the scale is the construct (e.g. IQ).
These measurement concepts can correspond to different mathematical estimation techniques.
Most PLS-based SEM use formative scales.
Different software frameworks may use different terminology (e.g. SmartPLS, SEMinR)
SmartPLS
reflective constructs
formative constructs
PLSc
SEMinR
Composite Mode A (cor)
Composite Mode B (reg)
Reflective
For this exercise, we assume that all our constructs are type A composite constructs.
Items (manifest variables) are often called as the abbreviation and a number (e.g. PE1).
Constructs (latent variables) can have a single or multiple items.
The names should correspond to the column names in our data.
This model can be used for preregistration (e.g. using OSF).
Data collection is simplified using the prepared scales (survey, experiments).
Here, we collected data for you [1].
Two main questions:
Single function call in SEMinR
SEMinR uses Partial-Least Squares (PLS) estimation
Measurement model:
Structural model:
Are these values “significant”?
What is bootstrappingg?
Structural Equation Modeling in HCI Research using SEMinR